Abstract:Multiple linear regression is widely used in statistical analysis. Based on common tools of the multiple linear regression in big data research, especially in the research of hierarchical structure data, a partial regression coefficient model is proposed in this paper. The total partial regression coefficient is calculated by using each partial regression coefficient at the lower part and the hierarchical matrix between the lower part and upper part. This paper validates that the new model is equivalent to the common models of multiple linear regression from the theoretical research and the real data. The new method can effectively solve the problem of privacy data in privacy protection research. Meanwhile, the new model can realize the parallel computation, which improves the capability of big data processing.

Abstract:In the scenarios of diagnosing bearing faults for large rotary machinery, the ower identification accuracy existed in traditional fault identification technique. Hence, based on the frequency-domain analysis, a novel data mining framework called associated frequency patterns mining framework (AFPMF) was proposed in this paper, which was consisted of data pre-processing, associated frequency pattern mining process and fault status monitoring. In the data pre-processing of AFPMF, time window was adopted to divide the machinery vibration data stream into multiple sub-blocks, and then Fast Fourier Transform (FFT) was employed to make the data sub-blocks time-frequency transform for frequency feature extraction. The associated frequency pattern tree with sliding window was also used to build a compact tree for data mining. Finally, the potential fault status with the vibration frequency existed in the mining results was identified to realize the fault status monitoring. After the comparison of AFPMF and the traditional methods in the bearing fault diagnosis, the results show that AFPMF had higher identification accuracy than other traditional ones.

Abstract:It is one of the hot issues to motivate more users to participate in mobile crowd sensing tasks and provide high-quality data. As for that in many online incentive mechanisms,questions for the data quality provided by users and credit marks of the users are not paid enough attention,an online incentive mechanisms for sensory tasks and the evaluation of user credit model are proposed in this paper.Based on the reality and history credit of users, the credit updating algorithm model is established and the multi-stage online incentive mechanism QOM based on credit updating is designed.The simulation results show that the algorithm can help the platform obtain better utility and improve the employment efficiency.

Abstract:The traditional PNN neural network has strong fault tolerance, simple learning process and fast training speed. Based on the traditional PNN neural network, this paper uses LMS to optimize its heart sound classification, and then improve heart sound classification and prediction. accuracy. The LMS-PNN neural network algorithm framing the heart sound signal using the window function, using the double threshold method to determine the value of the data, using the LMS algorithm to debug the corresponding parameters, and saving the denoised data in mat format, extracting The short-time autocorrelation coefficients and short-term power spectral densities of each heart sound are used, and PNN neural network is used to extract 40,000 sample data for training, and each heart sound is graded and predicted. After inputting the training data from the mode layer of the PNN neural network, it can be obtained through simulation test that the prediction accuracy of the LMS-PNN neural network can reach more than 96%.

Abstract:Correctly identifying and classifying pulse signals of whales and active sonar or communication signals are very important for improving the stability and reliability of the marine passive acoustic monitoring and underwater sonar or underwater acoustic communication systems. In this paper, the representative Click signals of whales and three kinds of traditional sonar signals (CW, LFM, HFM) are selected as research objects, and a method for classifying Sperm Whale clicks and traditional sonar signals based on time-frequency features is proposed. Firstly, the denoising and automatic signal extraction of whale clicks are realized by using filtering, wavelet denoising and endpoint detection methods. Then, based on the short-time Fourier transform of the four types of signals, polynomials are used to fit the signal time-frequency contours, and the coefficients of the fitted polynomial are extracted as the time-frequency features of signals. Finally, the four types of signals are classified and identified using Back Propagation (BP) neural network and Support Vector Machine respectively. The classification results verify the effectiveness of the proposed algorithm and method.

Abstract:This paper addresses the relationship among the complex indicators of area control sectors using covariance matrix and hierarchical clustering method. Seven complexity indicators are built considering the structural and flow distribution features of area control sectors. Thirty-seven area control sectors in China are chosen to clarify the correlations among complexity indicators using visualized covariance matrix. It is found that the area of each sector varies greatly (coefficient of variation 194.75%) while flow distribution of different routes in sample sectors has little change (coefficient of variation 9.02%). The larger the sector range is, the more uniform the flow distribution is on each path (correlation coefficient 0.54~0.59); the larger the sector range is, the lower the sector flow is (correlation coefficient -0.06~-0.13). Single linkage method is introduced to cluster these complexity indicators. Clustering tree shows that static complexity indicators and dynamic complexity indicators can be well distinguished setting the distance between classes 0.70. It is found that static complexity indicators share more commonness than dynamic ones.

Abstract:In order to realize the automatic removing of impurities in silicon particles, an impurity identification algorithm was presented and an automated impurity removal device was designed. Firstly, the image of silicon particles is collected by the camera in real time, and it is preprocessed, morphologically operated, and edge detected by Visual C++ combined with OpenCV to identify and locate the impurity. Then, based on the RS-232 protocol, the communication with the single-chip microcomputer is realized, the position information of the impurity is transmitted, the action of the corresponding device is controlled. Finally the separation of the silicon and the impurity is accomplished. It has been verified that the system can effectively identify and remove impurities, and the removal rate reaches 95%, which meets the design requirements. The device adopts the " subregional removal algorithm " combined with the "high-pressure gas" method for impurity removal, and has the advantages of fast action response and high rejection precision, improving production efficiency and reducing production cost.

Abstract:As a class of rateless erasure codes, Luby transform (LT) codes are successfully applied to the reliable data transmission in wireless communication. Degree distribution is critical to the performance of LT codes. However, the classical Robust Soliton Distribution (RSD) is not suitable for short-length LT codes. To solve the problem, a method is proposed to optimize the degree distribution for LT codes over Binary Erasure Channel (BEC). Based on the properties of degree distribution, the Artificial Fish Swarm Algorithm (AFSA) is used to optimize the proportions of some important degrees in RSD. Simulation results show that compared to other optimized degree distributions and RSD, the new degree distribution reduces the overhead, and saves the average consuming time per encoding/decoding process.

Abstract:This paper presents a data mining method combining improved genetic algorithm and association rules. Firstly, the crossover operator and mutation operator of genetic algorithm are improved adaptively so that they can adjust adaptively according to the fitness value of function in the process of iteration. The improved adaptive genetic algorithm is integrated into association rules to make full use of the good global search ability of genetic algorithm and improve the mining efficiency of association rules dealing with mass data. In order to avoid useless rules and reduce the existence of irrelevance, intimacy is added to improve the reliability of association rules. The optimized algorithm is verified by analyzing traffic data on Hadoop big data platform. Compared with traditional methods, this method improves the convergence speed and robustness of the algorithm.

Abstract:Abstract: A semiconductor defect recognition method based on Bi-2DPCA and improved convolution neural network is proposed in order to solve the surface defect problem of semiconductor devices in packaging process. In view of overcoming the problem of low recognition accuracy caused by uneven samples, the training image is transformed to construct virtual samples, and then the image is compressed with Bi-2DPCA to extract the main features of the image. The improved AlexNet network is used for defect recognition and classification. In order to solve the problem of low adaptability of the model caused by the diversity and unpredictability of the diode plastic sealing surface, a normal random sampling layer is proposed, which is added to the convolution layer of the AlexNet network for lower sampling. At the same time, DropConnect is introduced in the full connection layer to improve the generalization performance of the network.Experiments show that the proposed algorithm has a higher recognition rate than the related algorithm, and has been verified on the SMA surface image set. At the same time, the algorithm has a good generalization performance.

Abstract:In the speaker verification (SV) task, score normalization can improve the system performance by adjusting the score distribution of each speaker to a similar distribution. In this paper, a large number of imposter scores for the target speakers are obtained from the development set firstly, then these scores are clustered by unsupervised clustering algorithm and the Gaussian mixture models (GMM) are used to fit the score distribution. The mean and standard deviation of Gaussian component with maximum mean value are used in the SV score normalization method. Experiments are carried out on the NIST SRE 2016 test set. Compared with conventional score normalization methods, the proposed method can effectively improve the system performance.

Abstract:In Zero-IF receiver, the radio quadrature demodulator has inherent I/Q gain and phase imbalance. In order to reduce its impact on the demodulation performance, a novel FM discriminator is proposed. This algorithm can track and compensate the I/Q imbalance correctly without modeling, training and parameter-calculating. Without additional hardware cost, the proposed discriminator can be easily implemented. Simulation has proved that the proposed discriminator can get same demodulation performance both with I/Q imbalance and ideal quadrature demodulator. Besides, hardware experiments has been done to demonstrate its effectiveness and reliability.

Abstract:The azimuth resolution of civil navigation radar depends on the size of antenna aperture. Large aperture antenna is restricted by many factors in engineering project that it is difficult to be widely used. For this problem a super-resolution method based on Lagrange theory is proposed in this paper. The convolutional echo is converted into the form of matrix-vector product by supplementing elements to azimuth signal vector. Combining singular value decomposition (SVD) to antenna pattern with BFGS algorithm a new observed model is established and Lagrange function is utilized to solve the new model for restoring original azimuth signals, by which azimuth super-resolution can be realized. Experimental performances manifest that the proposed method has a good super-resolution, and a higher signal reconstruction to error ratio (SRER) is obtained comparing with dual-logarithmic method while the signal to noise ratio (SNR) ranges from 10dB to 20dB.

Abstract:In order to improve the tracking accuracy and robustness of correlation filter based trackers, especially with challenging factors like occlusion, fast motion, background clutter and so forth, an improved Deep Features with Local Constrained Mask Correlation Filter Tracking Algorithm is proposed. Based on the classical correlation filter tracking algorithm, the learned binary matrix is proposed as a local constrained mask to achieve pruning the filter energy, which concentrates the template information and effectively alleviates the boundary effects caused by circular shifted training samples. Deep features are introduced in the process of feature extraction. By exploiting rotation, flipping, and Gaussian blur operations, the training sample set is expanded, which makes feature templates learn more target information. Compared the robustness of our algorithm with mainstream methods under distractions like occlusion, background clutter and illumination changes.

Abstract:For the problem of underdetermined blind source separation (UBSS), a method to enhance signal sparsity is proposed, and the density based spatial clustering of applications with noise (DBSCAN) combined with the clustering by fast search and find of density peaks (CFSFDP) is used to estimate the mixing matrix. Firstly, the time domain observed signals are transformed into sparse signals in the time-frequency domain, the single-source-point (SSP) detection is used to highlight the linear clustering characteristics, and the mirroring mapping is used to transform the linear clustering into compact clustering for density-based clustering analysis. Then, in the dense data heaps, the DBSCAN is used to search for high-density points and their corresponding neighborhoods to automatically find the number of clusters and the initial cluster centers. Finally, the number of clusters is used as the input parameter of CFSFDP, and the corresponding density peaks are searched by CFSFDP in the range of data clusters to achieve further correction of the cluster centers position. The above method not only improves the estimation accuracy of the underdetermined mixing matrix, but also provides a highly consistent estimator.

Abstract:Voice conversion usually refers to the process of modifying and transforming the personalized features of a person’s voice to make it sound like another person’s voice while keeping the linguistic information unchanged. In recent years, with the rapid development of information processing and machine learning, the technology of voice conversion has also achieved great progress. On the basis of introducing the basic concepts of voice conversion, the typical models and methods of voice conversion researched in recent years are summarized in this paper, the key technologies of voice conversion are reviewed, the main application scenarios of voice conversion technology are listed, and some technical problems still existing in voice conversion at present are briefly introduced. Finally, the prospects for the directions of the research and development of voice conversion are given.

Abstract:Photoacoustic imaging, as a new imaging technique with high optical contrast and great ultrasonic detection depth, is a breakthrough of the barriers that the resolution and imaging depth of the traditional optical imaging technique are mutually restricted. It has obtained the unprecedented rapid development. In addition, photoacoustic microscopy inherits the advantages of the photoacoustic imaging technology, and has realized high-contrast and high-resolution biological structure, molecular and functional imaging by using acoustic or optical focusing imaging mode, and it has potential application value in neurology, ophthalmology, vascular biology and dermatology. Here, the principle and classification of photoacoustic imaging technology are firstly presented. Then, focusing on the theme of photoacoustic microscopy, we review its novel scanning methods, focal depth extension techniques and biomedical applications in depth. Finally, the challenges of the development of photoacoustic microscopy are summarized, and the future development direction is also prospected.

Abstract:Aiming at the performance degradation of traditional DOA(direction of arrival) estimation algorithm under the condition of array model error, a DOA estimation algorithm based on ResNet (residual network) is proposed. According to the characteristics of data-driven neural network that it does not depend on array flow pattern, the proposed algorithm extracts features from generalized cross-correlation (GCC), and takes the extracted features as input of neural network deep classifier to classify signals. On the basis of to the classification results, the corresponding sub-interval data are selected for training and the non-linear mapping relationship between ResNet learning features and DOA estimation to form a data-driven robust DOA estimation system. The simulation and experimental results show that the proposed algorithm can effectively solve the problem that the traditional DOA algorithm cannot accurately obtain the DOA results under the condition of array model error.

Abstract:It is one of the hot issues to motivate more users to participate in mobile crowd sensing tasks and provide high-quality data. As for that in many online incentive mechanisms, questions for the data quality provided by users and credit marks of the users are not paid enough attention, an online incentive mechanisms for sensory tasks and the evaluation of user credit model are proposed in this paper. Based on the reality and history credit of users, the credit updating algorithm model is established and the multi-stage online incentive mechanism ROM based on credit updating is designed. The simulation results show that the algorithm can help the platform obtain better utility and improve the employment efficiency.

Abstract:To improve the tracking accuracy and robustness of correlation filter based trackers, especially with challenging factors such as occlusion, fast motion, background clutter and so forth, an improved correlation filter tracking algorithm based on deep features with local constrained mask is proposed. Based on discriminative correlation filter tracking algorithms, the learned binary matrix is proposed as a local constrained mask to achieve pruning the filter energy, which suppresses the response map generated by the template edge and the testing images. This allows the proposed method to expand the target search region and effectively alleviates the boundary effects caused by circular shifted training samples. Deep features are introduced in the process of feature extraction. By exploiting rotation, flipping, and Gaussian blur operations, the training sample set is expanded, which makes feature templates learn more target information. Compared the robustness of our algorithm with the mainstream methods under distractions like occlusion, background clutter and illumination changes.

Abstract:For the problem of underdetermined blind source separation (UBSS), a method to enhance signal sparsity is proposed, and the density based spatial clustering of applications with noise (DBSCAN) combined with the clustering by fast search and find of density peaks (CFSFDP) is used to estimate the mixing matrix. Firstly, the time domain observed signals are transformed into sparse signals in the time-frequency domain, the single-source-point (SSP) detection is used to highlight the linear clustering characteristics, and the mirroring mapping is used to transform the linear clustering into compact clustering for density-based clustering analysis. Then, in the dense data heaps, the DBSCAN is used to search for high-density points and their corresponding neighborhoods to automatically find the number of clusters and the initial cluster centers. Finally, the number of clusters is used as the input parameter of CFSFDP, and the corresponding density peaks are searched by CFSFDP in the range of data clusters to achieve further correction of the cluster centers position. The above method not only improves the estimation accuracy of the underdetermined mixing matrix, but also provides a highly consistent estimator.

Abstract:Traditional probability neural network (PNN) has strong fault tolerance, simple learning process and fast training speed. To improve the performance of the traditional PNN in heart sound classification, we adopt least mean square (LMS) method to implement the optimization, thereby increasing the accuracy of heart sound classification and prediction. The LMS-PNN algorithm frames the heart sound signal using the window function, uses the double threshold method to determine the value of the data, employs the LMS algorithm to debug the corresponding parameters, and saves the denoised data in the format of mat file. It extracts the short-time autocorrelation coefficients and short- time power spectral densities of each heart sound, and uses PNN to extract 40 000 sample data for training. Each heart sound is graded and predicted. After inputting the training data from the mode layer of the PNN algorithm, experimental data verification shows that the prediction accuracy of LMS-PNN can reach more than 96%.

Abstract:In the speaker verification (SV) task, score normalization can improve the system performance by adjusting the score distribution of each speaker to a similar distribution. Here, a large number of imposter scores for the target speakers are obtained from the development set firstly, then these scores are clustered by unsupervised clustering algorithm and the Gaussian mixture models (GMM) are used to fit the score distribution. The mean and standard deviation of Gaussian component with maximum mean value are used in the SV score normalization method. Experiments are conducted on the NIST SRE 2016 test set and results show that compared with the conventional score normalization methods, the proposed method can effectively improve the system performance.

Abstract:Correctly identifying and classifying pulse signals of whales and active sonar or communication signals are very important for improving the stability and reliability of the marine passive acoustic monitoring and underwater sonar or underwater acoustic communication systems. In this paper, the representative Click signals of whales and three kinds of traditional sonar signals are selected as research objects, and a method for classifying sperm whale clicks and traditional sonar signals based on time-frequency features is proposed. Firstly, the denoising and automatic signal extraction of whale clicks are realized by using filtering, wavelet denoising and endpoint detection methods. Then, based on the short-time Fourier transform of the four types of signals, polynomials are used to fit the signal time-frequency contours, and the coefficients of the fitted polynomial are extracted as the time-frequency features of signals. Finally, the four types of signals are classified and identified using back propagation (BP) neural network and support vector machine respectively. The classification results verify the effectiveness of the proposed algorithm and method.

Abstract:Through the complex network theory, screening and dimension reduction in accordance with the TCGA(The cancer genome atlas) gastric cancer data are dealt with. We selected 275 genes related to gastric cancer, 40 samples of patients with stage IIB of gastric cancer and 36 samples of patients with stage IIIA of gastric cancer. By analyzing the gene change rate of the gastric cancer IIB sample group and the gastric cancer IIIA sample group, the joint relationship between the nodes (genes) is established. Due to the above work, the gene expression network in the process of gastric cancer deterioration is constructed. The network is analyzed by making uses of comprehensive central indicators, and 17 genes with higher comprehensive index are screened out. Using the relevant theory of complex networks to divide the gastric cancer gene network into communities, it is found that all the genes with higher index of 17 comprehensive centers fall in a large connected sub-network.This topology is consistent with the key nodes of the gastric cancer gene expression network. Therefore we have verified our conclusions on the other hand.Through comprehensive analysis, the key genes in the process of gastric cancer deterioration are obtained, which provided an effective early warning signal for gastric cancer deterioration.

Abstract:A data mining method is presented by combining improved genetic algorithm (GA) and association rules. Firstly, the crossover operator and mutation operator of GA are improved adaptively so that they can adjust adaptively according to the fitness value of function in the process of iteration. The improved adaptive GA is integrated into association rules to make full use of the good global search ability of GA and improve the mining efficiency of association rules dealing with mass data. To avoid useless rules and reduce the existence of irrelevance, intimacy is added to improve the reliability of association rules. The optimized algorithm is verified by analyzing traffic data on Hadoop big data platform. Compared with traditional methods, this method improves the convergence speed and robustness of the algorithm.

Abstract:In the scenarios of diagnosing bearing faults for large rotary machinery, the traditional fault identification technique usually has the problem of low identification accuracy. Hence, based on the frequency-domain analysis, a novel data mining framework of frequency patterns mining framework (AFPMF) is proposed in this paper, which consists of data pre-processing, associated frequency pattern mining process and fault status monitoring. In the data pre-processing of AFPMF, the time window is adopted to divide the machinery vibration data stream into multiple sub-blocks, and then fast Fourier transform (FFT) is employed to make the data sub-blocks time-frequency transform for frequency feature extraction. The associated frequency pattern tree with sliding window is also used to build a compact tree for data mining. Finally, the potential fault is identified according to the vibration frequency existing in the mining results. Thus the fault status monitoring is realized. The comparison results of AFPMF and the traditional methods in the bearing fault diagnosis show that AFPMF has higher identification accuracy than other traditional ones.

Abstract:Multiple linear regression (MLR) is widely used in statistical analysis. Based on common tools of the multiple linear regression in big data research, especially in the research of hierarchical structure data, a partial regression coefficient model is proposed here. The total partial regression coefficient is calculated by using each partial regression coefficient at the lower part and the hierarchical matrix between the lower and upper parts. It is validated that the new model is equivalent to the common models of multiple linear regression by the theoretical research and the real data. The new method can effectively solve the problem of privacy data in privacy protection research. Moreover, the new model can realize the parallel computation, which improves the capability of big data processing.

Abstract:In Zero-IF receiver, the radio quadrature demodulator has inherent I/Q gain and phase imbalance. In order to reduce its impact on the demodulation performance, a novel FM discriminator is proposed. The algorithm can track and compensate the I/Q imbalance correctly without modeling, training and parameter-calculating. Without additional hardware cost, the proposed discriminator can be easily implemented. Simulation proves that the proposed discriminator can get same demodulation performance both with I/Q imbalance and ideal quadrature demodulator. Besides, hardware experiments are achieved to demonstrate its effectiveness and reliability.

Abstract:As a class of rateless erasure codes, Luby transform (LT) codes are successfully applied to the reliable data transmission in wireless communication. Degree distribution is critical to the performance of LT codes. However, the classical robust soliton distribution (RSD) is not suitable for short-length LT codes. To solve the problem, a method is proposed to optimize the degree distribution for LT codes over binary erasure channel (BEC). Based on the properties of degree distribution, the artificial fish swarm algorithm (AFSA) is used to optimize the proportions of some important degrees in RSD. Simulation results show that compared to other optimized degree distributions and RSD, the new degree distribution reduces the overhead, and saves the average consuming time per encoding/decoding process.

Abstract:This paper addresses the relationship among the complex indicators of area control sectors using covariance matrix and hierarchical clustering method. Seven complexity indicators are built considering the structural and flow distribution features of area control sectors. Thirty-seven area control sectors in China are chosen to clarify the correlations among complexity indicators using visualized covariance matrix. It is found that the area of each sector varies greatly (coefficient of variation is 194.75%), while flow distribution of different routes in sample sectors has little change (coefficient of variation is 9.02%). The larger the sector range is, the more uniform the flow distribution is on each path (correlation coefficient is 0.54—0.59); the larger the sector range is, the lower the sector flow is (correlation coefficient is -0.06—-0.13). Single linkage method is introduced to cluster these complexity indicators. Clustering tree shows that static complexity indicators and dynamic complexity indicators can be well distinguished with the distance between classes of 0.70. It is found that static complexity indicators share more commonness than dynamic ones.

Abstract:The azimuth resolution of civil navigation radar depends on the size of antenna aperture. Large aperture antenna is restricted by many factors in engineering project, and it is difficult to be widely used. For this problem a super-resolution method based on Lagrange theory is proposed in this paper. The convolutional echo is converted into the form of matrix-vector product by supplementing elements to azimuth signal vector. Combining singular value decomposition (SVD) to antenna pattern with BFGS algorithm a new observed model is established, and Lagrange function is utilized to solve the new model for restoring original azimuth signals, by which azimuth super-resolution can be realized. Experimental performances manifest that the proposed method has good super-resolution, and a higher signal reconstruction to error ratio (SRER) is obtained comparing with dual-logarithmic method while the signal to noise ratio (SNR) ranges from 10dB to 20 dB.

Abstract:A semiconductor defect recognition method based on Bi-2DPCA and improved convolution neural network is proposed to solve the surface defect problem of semiconductor devices in packaging process. In view of overcoming the problem of low recognition accuracy caused by uneven samples, the training image is transformed to construct virtual samples, and then the image is compressed with Bi-2DPCA to extract the main features of the image. The improved AlexNet network is used for defect recognition and classification. To solve the problem of low adaptability of the model caused by the diversity and unpredictability of the diode plastic sealing surface, a normal random sampling layer is proposed, which is added to the convolution layer of the AlexNet network for lower sampling. At the same time, DropConnect is introduced in the full connection layer to improve the generalization performance of the network.Experiments show that the proposed algorithm has a higher recognition rate than the related algorithm, and is verified on the SMA surface image set. At the same time, the algorithm has a good generalization performance.

Abstract:In order to realize the automatic removing of impurities in silicon particles, an impurity identification algorithm is presented and an automated impurity removal device is designed. Firstly, an image of silicon particles is collected by the camera in real time, and it is preprocessed, morphologically operated, and edge detected by Visual C++ combined with OpenCV to identify and locate the impurity. Then, based on the RS-232 protocol, the communication with the single-chip microcomputer is realized, the position information of the impurity is transmitted, and the action of the corresponding device is controlled. Finally, the separation of the silicon and the impurity is accomplished. It is verified that the system can effectively identify and remove impurities, and the removal rate reaches 95%, which meets the design requirements. The device adopts the “subregional removal algorithm” combined with the “high-pressure gas” method for impurity removal, and has the advantages of fast action response and high rejection precision, improving production efficiency and reducing production cost.